Skip to content

fix(interpreter): reject negative persistent file descriptors#1802

Merged
chaliy merged 1 commit into
mainfrom
2026-05-28-fix-negative-fd-variable-vulnerability
May 28, 2026
Merged

fix(interpreter): reject negative persistent file descriptors#1802
chaliy merged 1 commit into
mainfrom
2026-05-28-fix-negative-fd-variable-vulnerability

Conversation

@chaliy
Copy link
Copy Markdown
Contributor

@chaliy chaliy commented May 28, 2026

Motivation

  • A persistent FD-cap bypass was found where ensure_persistent_fd_capacity treated fd <= 2 as exempt, which incorrectly allowed negative fd values from brace-var redirections to create unbounded entries.
  • This could let untrusted scripts bypass max_file_descriptors and exhaust memory by inserting many distinct negative keys for coproc buffers or exec fd table.

Description

  • Validate and reject negative fds by returning an error for fd < 0 in ensure_persistent_fd_capacity.
  • Narrow the standard-fd exemption to only 0..=2 so only real standard descriptors are exempt.
  • Add regression tests fd_limit_rejects_negative_fd_var_output and fd_limit_rejects_negative_fd_var_input to verify v=-1; exec {v}>... and v=-1; exec {v}<... fail and do not bypass the limit.

Testing

  • Ran cargo test -p bashkit fd_limit_rejects_negative_fd_var -- --nocapture which compiled and executed the added tests.
  • The two new threat-model tests (fd_limit_rejects_negative_fd_var_output and fd_limit_rejects_negative_fd_var_input) passed.

Codex Task

@cloudflare-workers-and-pages
Copy link
Copy Markdown

cloudflare-workers-and-pages Bot commented May 28, 2026

Deploying with  Cloudflare Workers  Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status Name Latest Commit Preview URL Updated (UTC)
✅ Deployment successful!
View logs
bashkit 5280632 Commit Preview URL May 28 2026, 09:15 AM

@chaliy chaliy force-pushed the 2026-05-28-fix-negative-fd-variable-vulnerability branch from 3009a6c to 5280632 Compare May 28, 2026 09:14
@chaliy chaliy merged commit 11357aa into main May 28, 2026
34 checks passed
@chaliy chaliy deleted the 2026-05-28-fix-negative-fd-variable-vulnerability branch May 28, 2026 13:08
chaliy added a commit that referenced this pull request May 28, 2026
## Release v0.8.0

Minor release. New feature: VFS-backed Python `open()` support.

### Highlights

- **Python `open()` support** — VFS-backed `open()` / `Path.open()`
read, write, and append now work in the embedded Python builtin, so
LLM-generated `with open("/tmp/...")` scripts run instead of failing.
Host filesystem and network stay unavailable to Python
([#1800](#1800)).
- Further `rg` parity and hardening fixes (default type globs, JSON
context fanout cap, root-arg allocation) plus interpreter fixes for
variable attribute/nameref persistence and persistent file descriptor
validation.

### What's Changed

* ci: reclaim runner disk before disk-hungry scheduled jobs
([#1807](#1807)) by @chaliy
* fix(rg): align r and tf default type globs with ripgrep
([#1805](#1805)) by @chaliy
* fix(rg): cap JSON context event fanout
([#1804](#1804)) by @chaliy
* fix(interpreter): persist var attrs and namerefs across shell state
restore ([#1803](#1803)) by
@chaliy
* fix(interpreter): reject negative persistent file descriptors
([#1802](#1802)) by @chaliy
* fix(rg): avoid root arg string cloning across candidates
([#1801](#1801)) by @chaliy
* feat(python): support vfs-backed open
([#1800](#1800)) by @chaliy
* feat(site): add bashkit logo assets
([#1799](#1799)) by @chaliy
* fix(ci): bypass pnpm `--` separator that breaks napi build flag
forwarding ([#1798](#1798)) by
@chaliy
* fix(site): add homepage canonical link header
([#1797](#1797)) by @chaliy

**Full Changelog**:
v0.7.2...v0.8.0

### Version bump

`0.7.2` → `0.8.0` across `Cargo.toml`, `crates/bashkit-cli/Cargo.toml`,
`crates/bashkit-js/package.json`, and `Cargo.lock`.

### Publish-readiness report

- ✅ `cargo fmt --check` — clean
- ✅ `cargo clippy --workspace --all-targets --features
http_client,ssh,sqlite -- -D warnings` — clean
- ✅ `cargo test --workspace --lib --bins --tests --features
http_client,ssh,sqlite` — green (with pinned ripgrep 15.1.0, matching
CI's `RG_VERSION`)
- ✅ `cargo test --features python -p bashkit` — green
- ✅ `cargo publish --dry-run -p bashkit` — succeeds after replicating
the workflow's monty/python strip step (git-only `monty` dep is stripped
before publish, as in `publish.yml`)
- ✅ `bashkit-cli` packages cleanly; its dry-run's only remaining step is
resolving `bashkit ^0.8.0` from crates.io, which the workflow satisfies
by publishing `bashkit` first (30s index wait) then `bashkit-cli`
- ✅ Version sync: all manifests read `0.8.0`; crates.io latest is
`0.7.2`, so `0.8.0` is a strict increment on every registry (npm/PyPI
versions are derived from the same pipeline)

Notes: in this sandbox the `rg` differential case and
`ssh_supabase_connects` test initially failed only because of
environment (preinstalled ripgrep 14.1.0 vs CI's pinned 15.1.0, and no
egress to the external `supabase.sh` SSH host); both are environmental,
not code regressions.


---
_Generated by [Claude
Code](https://claude.ai/code/session_01URKhGRWFuMgjHQR1YhyNHN)_
chaliy added a commit that referenced this pull request May 30, 2026
Reject negative fd values in ensure_persistent_fd_capacity and narrow the standard-fd exemption from `fd <= 2` to `0..=2` so attacker-supplied negative fd vars cannot bypass max_file_descriptors. Adds TM-DOS-063 regression tests.
chaliy added a commit that referenced this pull request May 30, 2026
## Release v0.8.0

Minor release. New feature: VFS-backed Python `open()` support.

### Highlights

- **Python `open()` support** — VFS-backed `open()` / `Path.open()`
read, write, and append now work in the embedded Python builtin, so
LLM-generated `with open("/tmp/...")` scripts run instead of failing.
Host filesystem and network stay unavailable to Python
([#1800](#1800)).
- Further `rg` parity and hardening fixes (default type globs, JSON
context fanout cap, root-arg allocation) plus interpreter fixes for
variable attribute/nameref persistence and persistent file descriptor
validation.

### What's Changed

* ci: reclaim runner disk before disk-hungry scheduled jobs
([#1807](#1807)) by @chaliy
* fix(rg): align r and tf default type globs with ripgrep
([#1805](#1805)) by @chaliy
* fix(rg): cap JSON context event fanout
([#1804](#1804)) by @chaliy
* fix(interpreter): persist var attrs and namerefs across shell state
restore ([#1803](#1803)) by
@chaliy
* fix(interpreter): reject negative persistent file descriptors
([#1802](#1802)) by @chaliy
* fix(rg): avoid root arg string cloning across candidates
([#1801](#1801)) by @chaliy
* feat(python): support vfs-backed open
([#1800](#1800)) by @chaliy
* feat(site): add bashkit logo assets
([#1799](#1799)) by @chaliy
* fix(ci): bypass pnpm `--` separator that breaks napi build flag
forwarding ([#1798](#1798)) by
@chaliy
* fix(site): add homepage canonical link header
([#1797](#1797)) by @chaliy

**Full Changelog**:
v0.7.2...v0.8.0

### Version bump

`0.7.2` → `0.8.0` across `Cargo.toml`, `crates/bashkit-cli/Cargo.toml`,
`crates/bashkit-js/package.json`, and `Cargo.lock`.

### Publish-readiness report

- ✅ `cargo fmt --check` — clean
- ✅ `cargo clippy --workspace --all-targets --features
http_client,ssh,sqlite -- -D warnings` — clean
- ✅ `cargo test --workspace --lib --bins --tests --features
http_client,ssh,sqlite` — green (with pinned ripgrep 15.1.0, matching
CI's `RG_VERSION`)
- ✅ `cargo test --features python -p bashkit` — green
- ✅ `cargo publish --dry-run -p bashkit` — succeeds after replicating
the workflow's monty/python strip step (git-only `monty` dep is stripped
before publish, as in `publish.yml`)
- ✅ `bashkit-cli` packages cleanly; its dry-run's only remaining step is
resolving `bashkit ^0.8.0` from crates.io, which the workflow satisfies
by publishing `bashkit` first (30s index wait) then `bashkit-cli`
- ✅ Version sync: all manifests read `0.8.0`; crates.io latest is
`0.7.2`, so `0.8.0` is a strict increment on every registry (npm/PyPI
versions are derived from the same pipeline)

Notes: in this sandbox the `rg` differential case and
`ssh_supabase_connects` test initially failed only because of
environment (preinstalled ripgrep 14.1.0 vs CI's pinned 15.1.0, and no
egress to the external `supabase.sh` SSH host); both are environmental,
not code regressions.


---
_Generated by [Claude
Code](https://claude.ai/code/session_01URKhGRWFuMgjHQR1YhyNHN)_
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant